cortico 0.1.2 → 0.1.3

package/src/providers/console/settings.ts

@@ -15,19 +15,19 @@ import { existsSync, mkdirSync, readdirSync, rmSync, writeFileSync } from 'node:
 import { join } from 'node:path';
 import { updateJsonObject } from '../../config-file.ts';
 import type { Language } from '../../core/language.ts';
-import type { ConsoleLamp, ConsolePageContribution } from '../../web/shared/console-protocol.ts';
+import { isConnectionEditorBlock, type ConsoleLamp, type ConsolePageContribution } from '../../web/shared/console-protocol.ts';
 import type { ConsolePageSource } from '../../web/console-pages.ts';
 import { providerModules, type ProviderRegistry } from '../registry.ts';
 import type { ProviderAvailability, ProviderModule } from '../base.ts';
 import { endpointAvailability, validateEntry } from '../configuration.ts';
 import { quotePrices, validatePrices, type PriceDefinition } from '../pricebook.ts';
-import { GenerationError } from '../../core/generation.ts';
+import { GenerationError, type ResponseClient } from '../../core/generation.ts';
 import { readTextFile } from '../../core/util.ts';
 import { responseRequest } from '../../protocol/open-responses/context-helpers.ts';
 import { record } from '../../protocol/open-responses/context.ts';
 import { text } from './strings.ts';
 import { PROBE_MAX_OUTPUT_TOKENS, type ProviderConsoleHost } from './types.ts';
-import { connectionGroup } from './config.ts';
+import { connectionBlocks, connectionGroup } from './config.ts';
 
 export type SecretStatus = 'env' | 'file' | 'none';
 
@@ -199,18 +199,24 @@ export class ProviderSettings {
     return readdirSync(dir).filter((file) => file !== 'config.json');
   }
 
-  private async probe(name: string, language: Language) {
+  private probe(name: string, language: Language) {
     const S = text(language);
     const entry = this.config.providers[name];
     if (!entry) throw new Error(S.unknownInstance);
     if (!entry.spec) throw new Error(S.specRequired);
+    return this.probeClient(this.registry.bind(name), entry.spec, language);
+  }
+
+  /** One diagnostic request through `client`; the receipt carries status, latency, usage and charges, or the failure and a hint. */
+  async probeClient(client: ResponseClient, spec: ModelSpec, language: Language) {
+    const S = text(language);
     const request = {
-      ...responseRequest(entry.spec, [record({ type: 'message', role: 'user', content: 'ping' })]),
-      max_output_tokens: Math.min(entry.spec.maxTokens ?? PROBE_MAX_OUTPUT_TOKENS, PROBE_MAX_OUTPUT_TOKENS),
+      ...responseRequest(spec, [record({ type: 'message', role: 'user', content: 'ping' })]),
+      max_output_tokens: Math.min(spec.maxTokens ?? PROBE_MAX_OUTPUT_TOKENS, PROBE_MAX_OUTPUT_TOKENS),
     };
     const started = Date.now();
     try {
-      const generation = await this.registry.bind(name).respond(request, { diagnostic: true, nativeSpec: entry.spec, role: 'probe' });
+      const generation = await client.respond(request, { diagnostic: true, nativeSpec: spec, role: 'probe' });
       const attempt = generation.attempts.at(-1);
       return {
         ok: true,
@@ -305,6 +311,12 @@ export class ProviderSettings {
       },
     };
     const extra = module.console?.(host) ?? {};
+    // A module that places none of the editor's blocks gets the default order, its own sections between model and pricing.
+    const declared = extra.panels ?? [];
+    const blocks = connectionBlocks(language);
+    const sections = declared.some((panel) => isConnectionEditorBlock(panel.builtin))
+      ? declared
+      : [blocks.endpoint, blocks.model, ...declared, blocks.pricing, blocks.protocol];
     return {
       ...extra,
       id: `llm:${module.id}`,
@@ -322,7 +334,7 @@ export class ProviderSettings {
           // 使用控制台内建端点面板，操作由下方 invoke 提供。
           builtin: 'llm-settings',
         },
-        ...(extra.panels ?? []),
+        ...sections,
       ],
       config: extra.config ?? entries.flatMap(
         ({ name, entry }) => module.config?.(name, entry, language) ?? [],

package/src/providers/console/strings.ts

@@ -14,6 +14,13 @@ const zh = {
   instancesBadge: '实例',
   settingsPanel: '实例与模型',
   settingsPanelDescription: "保存端点模型、价目及当前部署使用的端点。",
+  endpointBlock: '连接',
+  endpointBlockDescription: '模型服务的地址与 API Key。',
+  modelBlock: '模型与生成',
+  modelBlockDescription: '模型名与生成参数;模型列表从端点拉取。',
+  pricingBlock: '成本与计价',
+  protocolBlock: '高级协议',
+  protocolBlockDescription: '仅在使用自定义 API 网关或兼容服务时调整。',
   saved: '参数已保存；下一次请求生效。',
   unknownGroup: '未知 Provider 参数组',
   groupOutOfScope: 'Provider 参数组只能修改所属实例',
@@ -49,6 +56,13 @@ const en: typeof zh = {
   instancesBadge: 'Instances',
   settingsPanel: 'Instances and models',
   settingsPanelDescription: "Save the endpoint model, pricing and the active endpoint for this deployment.",
+  endpointBlock: 'Connection',
+  endpointBlockDescription: 'Provider URL and API Key.',
+  modelBlock: 'Model and generation',
+  modelBlockDescription: 'Model name and generation parameters; the model list comes from the endpoint.',
+  pricingBlock: 'Pricing',
+  protocolBlock: 'Advanced protocol',
+  protocolBlockDescription: 'Adjust only for custom API gateways or compatible services.',
   saved: 'Parameters saved; they apply from the next request.',
   unknownGroup: 'Unknown provider parameter group',
   groupOutOfScope: 'A provider parameter group can only modify its own instance',

package/src/providers/llamacpp/console/models-panel.ts

@@ -1,10 +1,13 @@
 /**
- * Model section of one endpoint: what llama-server knows, and pulling more.
- * The endpoint comes from `ctx.scope.instance`.
+ * Model section of one endpoint: what llama-server knows, a HuggingFace search that feeds the
+ * pull field, and `use` to put a served model into the model section. The endpoint comes from
+ * `ctx.scope.instance`.
  */
 import type { ConsolePanel, ConsolePanelContext } from '../../../web/shared/client-panel.ts';
+import type { HfFile, HfRepo } from '../huggingface.ts';
 import type { ModelsState } from './server.ts';
 import { panel } from '../strings.ts';
+import { HITS_PER_PAGE } from '../../../web/client/features/extensions/index.ts';
 
 const POLL_MS = 2_000;
 
@@ -22,8 +25,12 @@ export const modelsPanel: ConsolePanel = {
     const card = ui.sheet({ title: S.modelsTitle });
     const message = ui.msgline();
     root.append(card.el, message);
-    /** The pull input survives a re-render: the operator types while the poll redraws the card. */
+    /** Typed text and search results survive a re-render: the poll redraws the card around them. */
     let draft = '';
+    let query = '';
+    let repos: HfRepo[] | null = null;
+    let searchNote = '';
+    let chosen: { repo: string; files: HfFile[] } | null = null;
     let lastSnapshot = '';
 
     async function act(method: string, extra: Record<string, unknown> = {}): Promise<void> {
@@ -36,6 +43,76 @@ export const modelsPanel: ConsolePanel = {
       await load(true);
     }
 
+    async function search(text: string): Promise<void> {
+      const wanted = text.trim();
+      if (!wanted) return;
+      query = wanted;
+      chosen = null;
+      try {
+        // One page of results, the console's page size for search hits.
+        const result = await ctx.invoke<{ repos: HfRepo[] }>('search', [{ name, query: wanted, limit: HITS_PER_PAGE }]);
+        repos = result.repos;
+        searchNote = repos.length ? '' : S.noRepos;
+      } catch (error) {
+        repos = [];
+        searchNote = S.searchFailed(String(error));
+      }
+      await load(true);
+    }
+
+    async function showFiles(repo: string): Promise<void> {
+      try {
+        chosen = { repo, files: (await ctx.invoke<{ files: HfFile[] }>('files', [{ name, repo }])).files };
+      } catch (error) {
+        message.textContent = String(error);
+        return;
+      }
+      await load(true);
+    }
+
+    function renderSearch(body: HTMLElement): void {
+      const bar = ui.rowbar();
+      const input = ui.input({
+        placeholder: S.searchPlaceholder,
+        value: query,
+        onCommit: (value) => void search(value),
+      });
+      input.setAttribute('aria-label', S.search);
+      bar.append(input, ui.button(S.search, { onClick: () => void search(input.value) }));
+      body.append(ui.field(S.search, bar));
+      if (repos === null) return;
+      if (searchNote) body.append(ui.msgline(searchNote, repos.length === 0 && searchNote !== S.noRepos));
+      if (repos.length) {
+        const table = ui.table({ head: [S.repo, S.downloads, S.likes, S.updated, ''] });
+        for (const repo of repos) {
+          const open = ui.button(S.showFiles, { size: 'sm', onClick: () => void showFiles(repo.id) });
+          table.addRow([
+            { text: repo.id, cls: 'mono' },
+            ui.fmt.count(repo.downloads),
+            ui.fmt.count(repo.likes),
+            repo.updatedAt ? repo.updatedAt.slice(0, 10) : '',
+            open,
+          ]);
+        }
+        body.append(table.el);
+      }
+      if (!chosen) return;
+      body.append(ui.section(S.filesOf(chosen.repo)));
+      const files = ui.table({ head: [S.file, S.size, S.tag, ''] });
+      if (chosen.files.length === 0) files.clear(S.noFiles);
+      for (const file of chosen.files) {
+        const pull = ui.button(S.pull, { size: 'sm', variant: 'primary', onClick: () => void act('pull', { model: file.pull }) });
+        pull.title = file.pull;
+        files.addRow([
+          { text: file.name, cls: 'mono' },
+          file.bytes === null ? '' : bytes(file.bytes),
+          { text: file.tag ?? '', cls: 'mono' },
+          pull,
+        ]);
+      }
+      body.append(files.el);
+    }
+
     function render(state: ModelsState, body: HTMLElement): void {
       body.append(ui.kv([
         { k: S.cacheDir, v: state.cacheDir },
@@ -79,13 +156,18 @@ export const modelsPanel: ConsolePanel = {
           });
           status.append(progress.el);
         }
-        const actions = ui.h('div');
+        const actions = ui.rowbar();
         if (model.status === 'downloading')
           actions.append(ui.button(S.cancel, { size: 'sm', onClick: () => void act('cancel', { model: model.id }) }));
         else if (model.status === 'loaded' || model.status === 'sleeping')
           actions.append(ui.button(S.unload, { size: 'sm', onClick: () => void act('unload', { model: model.id }) }));
         else if (model.status === 'unloaded' || model.status === 'failed')
           actions.append(ui.button(S.load, { size: 'sm', onClick: () => void act('load', { model: model.id }) }));
+        if (model.status !== 'downloading') {
+          const use = ui.button(S.use, { size: 'sm', onClick: () => void act('use', { model: model.id }) });
+          use.title = S.useTitle;
+          actions.append(use);
+        }
         table.addRow([
           { text: model.id, cls: 'mono' },
           status,
@@ -95,6 +177,7 @@ export const modelsPanel: ConsolePanel = {
         ]);
       }
       body.append(table.el);
+      renderSearch(body);
     }
 
     async function load(force = false): Promise<void> {

package/src/providers/llamacpp/console/server.ts

@@ -5,11 +5,13 @@
  */
 import type { ConsolePageContribution } from '../../../web/shared/console-protocol.ts';
 import type { ProviderConsoleHost } from '../../console/types.ts';
+import { connectionBlocks } from '../../console/config.ts';
 import type { RouterCatalog, RouterModel } from '../catalog.ts';
 import { LAUNCH_DEFAULTS, PINNED_RELEASE, backendChoices, defaultBackend, llamacppOptions, type LaunchOptions } from '../options.ts';
 import type { LlamaRuntime } from '../runtime.ts';
 import type { RuntimeState } from '../runtime.ts';
 import { runtimeConfig } from '../config.ts';
+import { listGguf, searchGguf } from '../huggingface.ts';
 import { getByPath, type ConfigGroup, type ConfigValues } from '../../../core/config-schema.ts';
 import { text } from '../strings.ts';
 
@@ -49,11 +51,16 @@ export function llamacppConsole(host: ProviderConsoleHost): Partial<ConsolePageC
     if (!found) throw new Error(S.instanceNameRequired);
     return found.entry;
   };
+  const blocks = connectionBlocks(host.language);
   return {
     config: [],
     panels: [
-      { id: 'runtime', title: S.runtimePanel, description: S.runtimePanelDescription, slot: 'instance' },
-      { id: 'models', title: S.modelsPanel, description: S.modelsPanelDescription, slot: 'instance' },
+      { ...blocks.endpoint, title: S.endpointPanel, description: S.endpointPanelDescription },
+      { id: 'runtime', title: S.runtimePanel, description: S.runtimePanelDescription },
+      { id: 'models', title: S.modelsPanel, description: S.modelsPanelDescription },
+      blocks.model,
+      blocks.pricing,
+      blocks.protocol,
     ],
     invoke: async (panel, method, args) => {
       if (panel === 'runtime') {
@@ -140,6 +147,21 @@ export function llamacppConsole(host: ProviderConsoleHost): Partial<ConsolePageC
           await catalog.list(true);
           return { ok: true };
         }
+        if (method === 'search') {
+          if (typeof value.query !== 'string' || !value.query.trim()) throw new Error(S.queryRequired);
+          if (!Number.isInteger(value.limit) || (value.limit as number) < 1) throw new Error(S.limitRequired);
+          return { repos: await searchGguf(value.query.trim(), value.limit as number) };
+        }
+        if (method === 'files') {
+          if (typeof value.repo !== 'string' || !value.repo.trim()) throw new Error(S.repoRequired);
+          return { files: await listGguf(value.repo.trim()) };
+        }
+        if (method === 'use') {
+          if (typeof value.model !== 'string' || !value.model.trim()) throw new Error(S.modelIdRequired);
+          const entry = entryOf(name);
+          host.save(name, { ...entry, spec: { thinking: false, ...entry.spec, model: value.model.trim() } });
+          return { ok: true };
+        }
         if (host.editing) throw new Error(host.language === 'zh' ? '请先保存配置，再操作模型。' : 'Save configuration before model operations.');
         if (typeof value.model !== 'string' || !value.model.trim()) throw new Error(S.modelIdRequired);
         const model = value.model.trim();

package/src/providers/llamacpp/huggingface.ts

@@ -0,0 +1,93 @@
+/**
+ * HuggingFace lookups behind the pull field: repositories tagged `gguf`, and the GGUF files of
+ * one repository as `repo:tag` pull targets. The host is the one llama-server itself pulls from:
+ * `MODEL_ENDPOINT`, else `HF_ENDPOINT`, else huggingface.co.
+ */
+
+export interface HfRepo {
+  id: string;
+  downloads: number;
+  likes: number;
+  /** ISO timestamp of the last commit; null when the listing omits it. */
+  updatedAt: string | null;
+}
+
+export interface HfFile {
+  name: string;
+  /** Sum over shards; null when the listing carries no sizes. */
+  bytes: number | null;
+  /** Quantization tag read from the file name; null when the name carries none. */
+  tag: string | null;
+  /** What `POST /models` takes to pull this file: `repo:tag`, or the repository alone. */
+  pull: string;
+}
+
+export interface HfOptions {
+  fetchImpl?: typeof fetch;
+  env?: NodeJS.ProcessEnv;
+}
+
+const DEFAULT_ENDPOINT = 'https://huggingface.co';
+const SHARD_RE = /-(\d{5})-of-\d{5}\.gguf$/i;
+const QUANT_RE = /^(?:i?q\d[a-z0-9_]*|tq\d[a-z0-9_]*|bf16|f16|f32|f64|mxfp4[a-z0-9_]*)$/i;
+
+export function hfEndpoint(env: NodeJS.ProcessEnv = process.env): string {
+  return (env.MODEL_ENDPOINT || env.HF_ENDPOINT || DEFAULT_ENDPOINT).replace(/\/+$/, '');
+}
+
+async function getJson(url: string, fetchImpl: typeof fetch): Promise<unknown> {
+  const res = await fetchImpl(url, { headers: { Accept: 'application/json' } });
+  if (!res.ok) throw new Error(`GET ${url} ${res.status}`);
+  return res.json();
+}
+
+/** The `limit` most-downloaded matches; without one the listing returns up to a thousand rows. */
+export async function searchGguf(query: string, limit: number, options: HfOptions = {}): Promise<HfRepo[]> {
+  const params = new URLSearchParams({ search: query, filter: 'gguf', sort: 'downloads', direction: '-1', limit: String(limit) });
+  // The plain listing omits lastModified; `expand` names every field wanted back.
+  for (const field of ['downloads', 'likes', 'lastModified']) params.append('expand[]', field);
+  const rows = await getJson(`${hfEndpoint(options.env)}/api/models?${params}`, options.fetchImpl ?? fetch);
+  if (!Array.isArray(rows)) throw new Error('HuggingFace search returned no list');
+  const repos: HfRepo[] = [];
+  for (const row of rows as Array<Record<string, unknown>>) {
+    const id = typeof row.id === 'string' ? row.id : typeof row.modelId === 'string' ? row.modelId : null;
+    if (!id) continue;
+    repos.push({
+      id,
+      downloads: typeof row.downloads === 'number' ? row.downloads : 0,
+      likes: typeof row.likes === 'number' ? row.likes : 0,
+      updatedAt: typeof row.lastModified === 'string' ? row.lastModified : null,
+    });
+  }
+  return repos;
+}
+
+/** The quantization tag in a GGUF file name: the last `-` or `.` separated token that names one, with an `UD-` prefix kept. */
+export function quantTag(filename: string): string | null {
+  const stem = filename.replace(SHARD_RE, '.gguf').replace(/\.gguf$/i, '');
+  const tokens = stem.split(/[-.]/);
+  for (let i = tokens.length - 1; i >= 0; i--) {
+    if (!QUANT_RE.test(tokens[i])) continue;
+    return (tokens[i - 1]?.toUpperCase() === 'UD' ? 'UD-' : '') + tokens[i];
+  }
+  return null;
+}
+
+/** GGUF files of a repository, shards folded into one entry, projector files (`mmproj-*`) left out. */
+export async function listGguf(repo: string, options: HfOptions = {}): Promise<HfFile[]> {
+  const info = await getJson(`${hfEndpoint(options.env)}/api/models/${repo.split('/').map(encodeURIComponent).join('/')}?blobs=true`, options.fetchImpl ?? fetch) as { siblings?: unknown };
+  if (!Array.isArray(info.siblings)) throw new Error('HuggingFace repository listing carries no files');
+  const files = new Map<string, HfFile>();
+  for (const sibling of info.siblings as Array<Record<string, unknown>>) {
+    const path = typeof sibling.rfilename === 'string' ? sibling.rfilename : '';
+    if (!/\.gguf$/i.test(path) || /(^|\/)mmproj/i.test(path)) continue;
+    const name = path.replace(SHARD_RE, '.gguf');
+    const lfs = sibling.lfs as { size?: unknown } | undefined;
+    const size = typeof sibling.size === 'number' ? sibling.size : typeof lfs?.size === 'number' ? lfs.size : null;
+    const entry = files.get(name);
+    if (entry) { if (size !== null && entry.bytes !== null) entry.bytes += size; continue; }
+    const tag = quantTag(name);
+    files.set(name, { name, bytes: size, tag, pull: tag ? `${repo}:${tag}` : repo });
+  }
+  return [...files.values()];
+}

package/src/providers/llamacpp/strings.ts

@@ -14,6 +14,8 @@ const zh = {
   notManaged: '这条端点没有开启托管',
   runtimeNotInstalled: '运行时还没装好',
   runtimeUnsupported: '本机平台没有这个后端的官方构建,填自备运行时目录',
+  endpointPanel: '服务地址',
+  endpointPanelDescription: 'llama-server 监听的地址;托管的进程按它启动。',
   runtimePanel: '运行时',
   runtimePanelDescription: 'llama.cpp 官方二进制的下载、安装与 llama-server 进程的启停。',
   modelsPanel: '模型',
@@ -22,6 +24,9 @@ const zh = {
   unknownMethod: '未知操作',
   instanceNameRequired: '需要端点名',
   modelIdRequired: '需要模型 id',
+  queryRequired: '需要搜索词',
+  limitRequired: '需要结果条数',
+  repoRequired: '需要仓库名',
   bodyRequired: '需要请求体',
   serverUnreachable: "llama-server 不可达；请检查端点地址与服务状态。",
   externalServer: '端点已有服务在跑(外部启动),不接管',
@@ -51,6 +56,8 @@ const en: typeof zh = {
   notManaged: 'This endpoint is not managed',
   runtimeNotInstalled: 'The runtime is not installed yet',
   runtimeUnsupported: 'No official build for this backend on this platform; set an own runtime directory',
+  endpointPanel: 'Server address',
+  endpointPanelDescription: 'The address llama-server listens on; a hosted process starts on it.',
   runtimePanel: 'Runtime',
   runtimePanelDescription: 'Download and install the official llama.cpp binaries; start and stop the llama-server process.',
   modelsPanel: 'Models',
@@ -59,6 +66,9 @@ const en: typeof zh = {
   unknownMethod: 'Unknown action',
   instanceNameRequired: 'An endpoint name is required',
   modelIdRequired: 'A model id is required',
+  queryRequired: 'A search query is required',
+  limitRequired: 'A result count is required',
+  repoRequired: 'A repository is required',
   bodyRequired: 'A request body is required',
   serverUnreachable: "llama-server is unreachable; check the endpoint URL and server status.",
   externalServer: 'A server is already running at the endpoint (started externally); not taking over',
@@ -139,6 +149,22 @@ const panelZh = {
   reload: '重扫',
   pull: '拉取',
   pullPlaceholder: 'HuggingFace 仓库,如 ggml-org/Qwen2.5-Coder-7B-Instruct-GGUF',
+  use: '选用',
+  useTitle: '把这个模型填进「模型与生成」',
+  search: '搜索',
+  searchPlaceholder: '搜 HuggingFace 上的 GGUF 仓库,如 qwen3',
+  searchFailed: (reason: string) => `搜索失败: ${reason}`,
+  noRepos: '没有匹配的仓库。',
+  repo: '仓库',
+  downloads: '下载',
+  likes: '点赞',
+  updated: '更新',
+  showFiles: '看文件',
+  filesOf: (repo: string) => `${repo} 里的 GGUF 文件`,
+  noFiles: '这个仓库里没有 GGUF 文件。',
+  file: '文件',
+  size: '大小',
+  tag: '标签',
   localDir: '本机 GGUF 目录',
   cacheDir: '缓存目录',
   noModels: '暂无可用的 GGUF 模型：可通过上方输入拉取，或将 .gguf 文件放入本机目录后点击重扫。',
@@ -213,6 +239,22 @@ const panelEn: typeof panelZh = {
   reload: 'Rescan',
   pull: 'Pull',
   pullPlaceholder: 'HuggingFace repo, e.g. ggml-org/Qwen2.5-Coder-7B-Instruct-GGUF',
+  use: 'Use',
+  useTitle: 'Put this model into the model section',
+  search: 'Search',
+  searchPlaceholder: 'Search GGUF repositories on HuggingFace, e.g. qwen3',
+  searchFailed: (reason: string) => `Search failed: ${reason}`,
+  noRepos: 'No matching repositories.',
+  repo: 'Repository',
+  downloads: 'Downloads',
+  likes: 'Likes',
+  updated: 'Updated',
+  showFiles: 'Files',
+  filesOf: (repo: string) => `GGUF files in ${repo}`,
+  noFiles: 'This repository has no GGUF files.',
+  file: 'File',
+  size: 'Size',
+  tag: 'Tag',
   localDir: 'Local GGUF directory',
   cacheDir: 'Cache directory',
   noModels: 'No models yet: pull one above, or drop GGUF files into the local directory and rescan.',

package/src/providers/name.ts

@@ -6,3 +6,8 @@ export function validateProviderName(name: string): string | null {
     return 'This name is reserved by Windows.';
   return null;
 }
+
+/** 密钥变量名缺省时的默认名:按端点名派生,`-` 折成 `_`,与 validateEntry 的环境变量名格式一致。 */
+export function defaultSecretName(name: string): string {
+  return `CORTICO_KEY_${name.toUpperCase().replace(/-/g, '_')}`;
+}

package/src/providers/openai-responses-compat/console/server.ts

@@ -6,6 +6,7 @@
  */
 import type { ConsolePageContribution } from '../../../web/shared/console-protocol.ts';
 import { PROBE_MAX_OUTPUT_TOKENS, type ProviderConsoleHost } from '../../console/types.ts';
+import { connectionBlocks } from '../../console/config.ts';
 import { getByPath, type ConfigGroup, type ConfigValues } from '../../../core/config-schema.ts';
 import type { LLMProviderEntry, ModelSpec } from '../../../core/types.ts';
 import { GenerationError, type Generation, type ResponseClient } from '../../../core/generation.ts';
@@ -103,9 +104,16 @@ export function compatConsole(host: ProviderConsoleHost): Partial<ConsolePageCon
     if (verdict) host.save(name, { ...entry, options: { ...entry.options, reasoningReplay: verdict } });
     return { verdict, bare, withReasoning };
   };
+  const blocks = connectionBlocks(host.language);
   return {
     config: [],
-    panels: [{ id: 'reasoning', title: S.reasoningPanel, description: S.reasoningPanelDescription, slot: 'instance' }],
+    panels: [
+      blocks.endpoint,
+      blocks.model,
+      { id: 'reasoning', title: S.reasoningPanel, description: S.reasoningPanelDescription },
+      blocks.pricing,
+      blocks.protocol,
+    ],
     invoke: async (panel, method, args) => {
       if (panel !== 'reasoning') throw new Error(S.unknownPanel);
       const name = body(args);

package/src/providers/openai-responses-compat/native.ts

@@ -28,8 +28,8 @@ export interface ResponsesProviderOptions {
 
 /**
  * Send reasoning only when an effort is supplied; otherwise leave it to the endpoint.
- * Requests set store=false and request encrypted reasoning content for local stateless replay;
- * how recorded reasoning is replayed follows `reasoningReplay`. Endpoint extraBody fields are merged last.
+ * Requests set store=false, so replay is local and follows `reasoningReplay`; encrypted replay asks
+ * for the signed blocks back through `include`. Endpoint extraBody fields are merged last.
  */
 export function buildResponsesBody(
   request: Request,
@@ -41,10 +41,15 @@ export function buildResponsesBody(
   const body: Item = {
     ...request,
     input: replay.input,
-    include: [...new Set([...(request.include ?? []), 'reasoning.encrypted_content'])],
     store: false,
     stream: Boolean(options.onEvent),
   };
+  // 加密形态靠 include 把签名推理块取回来,明文形态用不上它;端点不支持 include 时整条请求被拒,
+  // 所以只在用得上的形态带。
+  const include = new Set<string>(request.include ?? []);
+  if (input.reasoningReplay !== 'plaintext') include.add('reasoning.encrypted_content');
+  if (include.size) body.include = [...include];
+  else delete body.include;
   delete body.previous_response_id;
   if (replay.instructions !== undefined) body.instructions = replay.instructions;
   else delete body.instructions;

package/src/providers/pricebook.ts

@@ -2,16 +2,15 @@ import { createHash } from 'node:crypto';
 import type { Request } from '../protocol/open-responses/index.ts';
 import type { LLMProviderEntry } from '../core/types.ts';
 import type { Language } from '../core/language.ts';
-import type { PriceSnapshot, PriceRule } from '../core/generation.ts';
+import type { PriceSnapshot, PriceRule, PriceTable, PriceWindow } from '../core/generation.ts';
 import { text } from './strings.ts';
 
-export interface PriceDefinition {
+export interface PriceDefinition extends PriceTable {
   models: string[];
   currency: string;
   basis: 'marginal' | 'equivalent';
-  rules: PriceRule[];
-  inputBands?: Array<{ from: number; rules: PriceRule[] }>;
-  serviceTiers?: Record<string, { rules: PriceRule[]; inputBands?: Array<{ from: number; rules: PriceRule[] }> }>;
+  /** A module declares its own time-of-day windows; `validatePrices` does not take them from an endpoint. */
+  timeWindows?: PriceWindow[];
   source: string;
 }
 export interface QuoteTime { startedAt: string; requestedServiceTier: string | null; }
@@ -60,6 +59,7 @@ export function validatePrices(value: unknown, language: Language = 'zh'): Price
     if (typeof raw.currency !== 'string' || !raw.currency.trim()) throw new Error(S.currencyRequired);
     if (raw.basis !== 'marginal' && raw.basis !== 'equivalent') throw new Error(S.basisValue);
     if (typeof raw.source !== 'string' || !raw.source.trim()) throw new Error(S.sourceRequired);
+    if (raw.timeWindows !== undefined) throw new Error(S.timeWindowsUnsupported);
     const tiers: PriceDefinition['serviceTiers'] = {};
     if (raw.serviceTiers !== undefined) {
       if (!raw.serviceTiers || typeof raw.serviceTiers !== 'object' || Array.isArray(raw.serviceTiers)) throw new Error(S.tiersObject);

package/src/providers/registry.ts

@@ -1,4 +1,4 @@
-import { existsSync, readdirSync } from 'node:fs';
+import { existsSync, readdirSync, readFileSync } from 'node:fs';
 import { join } from 'node:path';
 import { URL, fileURLToPath, pathToFileURL } from 'node:url';
 import type { LLMProviderEntry } from '../core/types.ts';
@@ -56,6 +56,8 @@ export class ProviderRegistry {
     private readonly entries: () => Record<string, LLMProviderEntry>,
     private readonly host: ProviderHostBase,
     private readonly modules: readonly ProviderModule[] = providerModules,
+    /** Secrets read before the process environment and the endpoint `.env`; preview registries carry the typed, unsaved key here. */
+    private readonly secretOverrides: Readonly<Record<string, string>> = {},
   ) {}
 
   private module(kind: string): ProviderModule {
@@ -64,9 +66,16 @@ export class ProviderRegistry {
     return module;
   }
 
-  /** Preview instances use a detached entry and never enter the live instance cache. */
+  /**
+   * A registry over one detached entry. Its instances never enter the live cache; `secrets`
+   * are the values the console holds but has not written to the endpoint's `.env`.
+   */
+  previewRegistry(name: string, entry: LLMProviderEntry, secrets: Readonly<Record<string, string>> = {}): ProviderRegistry {
+    return new ProviderRegistry(() => ({ [name]: entry }), this.host, this.modules, secrets);
+  }
+
   preview(name: string, entry: LLMProviderEntry): ProviderInstance {
-    return new ProviderRegistry(() => ({ [name]: entry }), this.host, this.modules).resolve(name);
+    return this.previewRegistry(name, entry).resolve(name);
   }
 
   resolve(name: string): ProviderInstance {
@@ -75,16 +84,20 @@ export class ProviderRegistry {
     const module = this.module(raw.kind);
     const entry = module.normalize?.(structuredClone(raw)) ?? structuredClone(raw);
     const { pricing: _pricing, spec: _spec, ...transportEntry } = entry;
-    const key = JSON.stringify(transportEntry);
+    // 密钥在实例创建时定格,`.env` 的内容指纹进缓存键:文件变更后下一次解析重建实例。
+    const stateDir = join(this.host.stateRoot, name);
+    const envFile = join(stateDir, '.env');
+    const fingerprint = existsSync(envFile) ? createHash('sha256').update(readFileSync(envFile)).digest('hex') : '';
+    const key = JSON.stringify([transportEntry, fingerprint]);
     const previous = this.instances.get(name);
     if (previous?.key === key) return previous.value;
     // 按端点名分岔的两样在这里填:模块自己的目录,与只读那个目录的密钥链。
-    const { stateRoot, ...base } = this.host;
-    const stateDir = join(stateRoot, name);
+    const { stateRoot: _stateRoot, ...base } = this.host;
+    const stored = secretReader(envFile);
     const value = module.create(name, entry, {
       ...base,
       stateDir,
-      secret: secretReader(join(stateDir, '.env')),
+      secret: (key) => this.secretOverrides[key] ?? stored(key),
       currentEntry: () =>
         module.normalize?.(structuredClone(this.entries()[name])) ?? this.entries()[name],
       resource: <T>(resource: string, create: () => T): T => {
@@ -124,11 +137,7 @@ export class ProviderRegistry {
     return client;
   }
 
-  /**
-   * Drop the cached instance so the next resolve rebuilds it; `host.resource` objects
-   * survive. Secrets are read once per instance, so a key written to the endpoint's `.env`
-   * takes effect only through this.
-   */
+  /** Drop the cached instance so the next resolve rebuilds it; `host.resource` objects survive. */
   invalidate(name: string): void {
     this.instances.delete(name);
   }

package/src/providers/strings.ts

@@ -31,6 +31,7 @@ const zh = {
   tiersObject: '服务档报价必须是对象',
   tierNameRequired: '服务档名不能为空',
   tierRules: '服务档报价需要规则对象',
+  timeWindowsUnsupported: '分时段价目由 provider 扩展在模块价目里声明,端点价目不接受 timeWindows',
   noModel: '未选模型',
   noSecret: (name: string) => `缺少密钥 ${name}`,
 };
@@ -65,6 +66,7 @@ const en: typeof zh = {
   tiersObject: 'Service tier pricing must be an object',
   tierNameRequired: 'Service tier name cannot be empty',
   tierRules: 'Service tier pricing needs a rules object',
+  timeWindowsUnsupported: 'Time-of-day pricing is declared by a provider extension in its module quote; endpoint pricing does not accept timeWindows',
   noModel: 'No model selected',
   noSecret: (name: string) => `Missing secret ${name}`,
 };